Académique Documents
Professionnel Documents
Culture Documents
Research Scholar, Swami Keshwanand Institute of Technology, Management & Gramothan, Jaipur, India
Associate professor, Swami Keshwanand Institute of Technology, Management & Gramothan, Jaipur, India
ABSTRACT
This paper presents a supervising learning approach using Multilayer Feed Forward Neural Network(MFFN) and
Radial Basis Fuction Neural Network(RBFN) to deal with fast and accurate static security assessment (SSA) and contingency
analysis of a large electric power systems. The degree of severity of contingencies is measured by two scalar performance
indices (PIs): Voltage-reactive power performance index, PIVQ and line MVA performance index, PIMVA. For each (N-1)
feature selection technique has been utilized to identify the inputs for the MFFN and RBFN. The proposed method has been
applied on an IEEE 39-bus New England test system at different operating conditions comparing to single line outage and the
results demonstrate the suitability of the methodology for on-line power system security assessment at Energy Management
Center. The performace of the proposed ANN models is compared withNewton Raphson (NR) method and the results shows
that the proposed model is effective and reliable in terms of static security assessment of power systems.
KEYWORDS: PerformanceIndex, Static Security Assessment, Contingency Analysis, Supervised Learning, Multilayer Feed
Forward Neural Network, Radial Basis function Network
Received: Mar 28, 2016; Accepted: Apr 15, 2016; Published: Apr 25, 2016; Paper Id.: IJEEERJUN201602
INTRODUCTION
Security assessment of a power systemplays an important role for on-line applicationsat Energy Management
Centerwhich employs to examine the steady state performance of a power system after contingency. Nowdays, electric
power system move towards a new environment that is deregulation which has forced modern electric utilities to operate
their systems under highly stressed conditions closer to their security limits. Therefore, the system operators needs to
develop quick and more precise ranking methods for analyzing the power system security violations and severity level
of contingencies to keep the power system in safe operating limits. As indicated by the fast development of Machine
Learning (ML) applications in power system area, this paper presents a review of Artificial Neural Network (ANN) in
Static Security Assessment (SSA).
Many research papers have discussed different methods to simulate and rank the contingencies for instance,
automatic contingency selection based on a pattern analysis as proposed by Rodrigues [Rodrigues, 1999]. This method
is capable to identify the potential harmful contingencies.On the other hand, Dynamic security assessment (DSA)
enables to determine which contingencies may cause power system operating limit violations or system instability.So
far, the only Transient Energy Function (TEF) analytical methods are suggested for dynamic security assessment [2] and
the use of Artificial Intelligence including Neural Networks and Expert Systems [3]. In [4] Transient stability
assessment of power system using Probabilistic Neural Network (PNN) and Least Squares Support Vector Machine
www.tjprc.org
editor@tjprc.org
Original Article
contingency, thePerformance Index (PI) is computed using the Newton Raphson (NR) method. A correlation coefficient
10
(LS-SVM) was presented.For a given severity of contingencies any on-line TSA tool must provide a fast and accurate
stability evaluation and system security analysis under random load perturbations.System security can be assessed using
contingency analysis. Contingency analysis identified those cases which are harmful to the system and ranked them
according to their severity is referred to as contingency selection.Ranking methods employs two scalar performance index
(PI) to measure the severity of each single line outage [5-7].In [8] a cascade neural network-based approach is proposed for
fast line flow contingency selection and ranking. The developed cascade neural network is a combination of a ranking
module and a filter module. Generally PI-based methods are difficult for online applications because of high computational
time needed.
Recently, artificial neural networks have shown a great promise for contingency screening and ranking due to
their ability to accurate input-output mappings. To estimate the severity of contingencies a pattern recognition technique is
proposed in [9-11].Over the past few years, several approaches have been proposed for static security assessmentusing
artificial neural networks (ANN) [12][14]. In [15] author employs ANN for online contingency screening and ranking and
found that ANN is effective in terms of accuracy and speed. In this paper,two four-layered Multilayer Feed Forward
Neural Network(MFFN) and Radial Basis Function Neural Network(RBFN) are proposedwhich predict the degree of
severity of contingencies.The study involves RBFN is more accurate and faster than MFFN as RBF networks reduce the
computation time required for training. Further due to minimum training error, the optimal learning can be achieved by
RBF networks. The proposed ANN models are trained with Resilient back propagation algorithm for inputoutput
mapping [16].
PI
MVA
Where
Nl
W Li
l =1 M
post
l
post
S lmax
S l
2n
(1)
max
l
the number of
the real non-negative weighting factor (=1). n is the integer exponent. In this paper the value of
PI
VQ
NB
= W Vi
M
i =1
Where
Lim
i
sp
V i V i
Lim
Vi
Vi
11
NG
+ W Gi
i =1 M
Qi
max
Q i
2n
(2)
sp
i
= rated voltage
max
= upper
limit for reactive power generation at bus i; NG the number of generating units, NB number of buses, n is the integer
exponent where n=4;
Gi
= real non-negative weighting factor(=1). According to the PIs value, three different security
states have been considered for the contingencies which are mentioned as; Secure state if PI < 0.2, Critical state if 0.2 < PI
< 0.8, Insecure state if PI > 0.8.
PROPOSED METHODOLOGY
Proposed ANN Methodology for Performance Indices
The proposed ANN model selected for on-line security assessment is a multilayer layer feed forward neural
network (MFFN) and Radial Basis Function Neural Network (RBFN) trained with Resilient back propagation algorithm.To
avoid misclassification, separate ranking is obtained for PIVQ and PIMVA using two MFFNs and RBFN as shown in the
block diagram of the proposed model in Figure 1(a) and Figure 1(b). For each contingency, the performance indexes are
calculated by off-line Newton Raphson method.
MFF N-I
YI=[PG10,QG1,QG2,
QG3,QG4,QG5,QD14,
QG7,QG8,QG9,QG10]
YO=PIVQ
Input
layer
Hidden
layer
Output
layer
MFFN-II
YO=PIMVA
Input
layer
Hidden
layer
Output
layer
editor@tjprc.org
12
Input X1
Weights
Radial bias
function
f1
W1
W0
Linear
weights
Input X2
W2
Output Y
f2
WK
Output Layer
fm
Hidden
Layer
Input Xk
Input Layer
i
i ( X ) = exp 2 2
i
(3)
(X ) =
Where,
i =1
ki
w
ki
(X ) +
(4)
ko
ko
the basis function. In this paper, RBFNgives faster convergence than the conventional MFFN.
13
Feature Selection
One of the key issues in ANN-based approach is the proper selection of inputs to the neural networks training. In
this paper correlation coefficient technique has been used to select the appropriate training features for the MFNN and
RBFN. The correlation coefficient between the jth and kth variable is calculated using (3) as
=
jk
{X X } E{X } E {X }
j
j, k = 1,2,3,4,,n
(5)
The variables that are less correlated are selected as features for the MFNN and RBFN. In this paper precontingent real and reactive power generation at the generator buses (PG, QG), reactive power loads (QD) are selected as
the input features for the proposed ANN training. Thus, 11 features (out of 62 features) are selected for the training of
MFNN and RBFN.
Data Generation, Training, and Testing
In this paper, single line contingency is considered. The load patterns are generated by changing the loads
randomly at each bus in the range of 80 and 143 % of their base case values. For each load pattern (N-1) contingency is
simulated by Newton Raphson method. For each contingency, the Performance Index (PI) is calculated using (1) and (2).
Nearly 11000 training patterns are generated for the proposed ANN model. The input features selected by correlation
coefficient method is normalized in the range of 0.1 and 0.9 for each load pattern. Once the training of ANNs is
accomplished, the trained network is evaluated by test data. The contingencies are ranked in terms of severity based on the
performance index PI.
editor@tjprc.org
14
MLFFN
1000
0
1x e-10
0.05
0.90
-
RBFN
0.1
Line
No.
Out
Age
No.
PIVQ
MFN
RBF
N-1
142.5 12-13
8069
0.9000 0.8856 0.8950
100
4-14
6072
0.2255 0.2628 0.2210
140
25-26
5206
0.3594 0.3031 0.3480
105
6-7
6699
0.3080 0.2925 0.2993
115
14-15
8028
0.1786 0.1878 0.1710
120
23-24
9062
0.1896 0.1908 0.1808
132.5
9-39
3141
0.5076 0.4806 0.4980
Class I and Class II (critical); Class III (non-critical).
NR
Class
(PIVQ)
NR
III
II
II
II
I
I
II
MFN
N-1
III
II
II
II
I
I
II
Class
(PIMVA)
PIMVA
NR
0.2296
0.1235
0.2056
0.1743
0.1301
0.1428
0.2143
MFNN2
0.2205
0.1225
0.2068
0.1721
0.1327
0.1389
0.2221
NR
RBF
0.2270
0.1220
0.2040
0.1715
0.1298
0.1405
0.2118
II
I
II
I
I
I
II
MFN
N-2
II
I
II
I
I
I
II
11,000 load patterns were generated by varying the loads randomly at all the buses and generation in the range of
80 143% of their base case values. For 240 different loading conditions, 46 single line outages are simulated for each
loading, to obtain different operating conditions. Contingency analysis has been performed by utilizing the pre-contingency
data and the line performance indices for each load scenario and each outage at a time. A total of 8840 patterns have been
taken to analyze the performance of the proposed model and for remaining cases Newton Raphson (NR) failed to converge.
The test results of the proposed MFNN and RBF model for contingency screening and ranking is shown in Table 2. It is
observed that PI values obtained by the proposed MFNN and RBF model are close to desired values of PI obtained from
Newton Raphson method. It can be concluded from the table that for performance indices, PIVQ and PIMVA a separate
ranking must be done. For sample result for 132.5 % of base case, it is found that operating constraints are violated for
outage of line 939, resulting in system is insecure as both the PIs are in insecure classes. Such result is expected since bus
No. 39 is a generator bus and line outage 939 causes bus voltage limit violation, the reactive power generation limit
violation and the overloading of the transmission lines connected to bus No. 9. The outage of line 1213 for both 142.5%
of base case makes the system critical as again all the operational constraints are violated. Similar analysis can be drawn
for other line outages. The number of cases that belong to the Class (I to III) obtained from the proposed MFNN and RBF
model and their comparison with Newton Raphson method is shown in Table 2. Table 1 shows the parameters for the
proposed MFNN and RBF model. The PI values obtained from MFNN and RBFneural network are very close with PI
values obtained from Newton Raphson method.
CONCLUSIONS
In this paper power system static security assessment has been investigated. The test results presented on IEEE
39-bus system provides the following observations.
15
A new method has been developed for calculating voltage performance and line performance index for
contingency ranking which eliminates misranking and masking problems.
The performace of the proposed models is compared with Newton Raphson (NR) method and the results shows
that the RBF model is more effective and reliable in terms of static security assessment of power systems.
Training is very fast as the RBF network has the capability of handling large amount of data.
REFERENCES
1.
Rodrigues, M S, Souza, J C S, Do CouttoFilho, M B and Th. Schilling, M, Automatic Contingency Selection Based on a
Pattern Analysis Approach, Proceedings of IEEE International Conference on Electric Power Engineering, Power Tech
Budapest 1999, pp179.
2.
G.D. Irisari, G.C. Ejebe, J.G. Waight, W.F. Tinney, Efficient Solution for Equilibrium Point in Transient Energy Function
Analysis, IEEE Trans. on Power Systems, May 1994, pp. 693-699.
3.
EPRI Report, lW3 103-02: Dynamic Security Analysis-Feasibility Evaluation Report, April 1994.
4.
Noor W, Azah M, Aini H. Transient stability assessment of a power system using PNN and LS-SVM methods. J ApplSci 2007;
7(21):320816.
5.
Swarup KS, Sudhakar G. Neural network approach to contingency screening and ranking in power systems. Neuro computing
2006; 70:10518.
6.
Chen, Y.; Bose, A. Direct ranking for voltage contingency selection. IEEE Transactions on
7.
8.
Ghosh S, Chowdhury BH. Design of an artificial neural network for fast line flow contingency ranking. Electr Power Energy
Syst 1996; 18(5):2717.
9.
Singh R, Srivastava L. Line flow contingency selection and ranking using cascade neural network. Neurocomputing 2007;
70:264550.
10. K. L. Lo, L. J. Peng, J. F. Macqueen, A. O. Ekwue, and D. T. Y. Cheng, Fast real power contingency ranking using a counterpropagation network, IEEE Trans. Power Syst., vol. 13, pp. 12591264, Nov. 1998.
11. Swarup KS. Artificial neural network using pattern recognition for security assessment and analysis. Neurocomputing 2008;
71:98398.
12. Souza J, Filho M, Schilling M. Fast contingency selection through a pattern analysis approach. Electr Power Syst Res 2002;
62:139.
13. V. S. Vankayala and N. D. Rao, Artificial neural network and their application to power systemA bibliographical survey,
Elect. Power Syst. Res., vol. 28, pp. 6769, 1993.S. Saeh and A. Khairuddin, Static security assessment using artificial neural
network, in Proc. IEEE Int. Conf. Power and Energy, Dec. 2008, pp. 11721177.
14. T. S. Sidhu and C. Lan, Contingency screening for steady-state security analysis by using FFT and artificial neural
networks, IEEE Trans. Power Syst., vol. 15, no. 1, pp. 421426, Feb. 2000.
15. R. Fischl, Application of neural networks to power system security: Technology and trends, in Proc. IEEE World Congr.
Computational Intelligence, Jul. 1994, vol. 6, pp. 37193723.
www.tjprc.org
editor@tjprc.org
16